Model-agnostic and Scalable Counterfactual Explanations via Reinforcement Learning

Counterfactual instances are a powerful tool to obtain valuable insights into decision processes, describing the necessary minimal changes in the space to alter the prediction towards a desired target. Most previous require a separate, computationally expensive optimization procedure instance, making them impractical for both large amounts of data and-dimensional data. Moreover, these methods are often restricted to certain of machine learning models (e. g. differentiable or tree-based). In this work, we propose a deep reinforcement learning approach that the optimization procedure into an end-to-end learnable process, us to generate batches of counterfactual instances in a single forward. Our experiments on real-world data show that our method i) is-agnostic (does not assume differentiability), relying only on feedback model predictions; ii) allows for generating target-conditional instances; iii) allows for flexible feature range constraints numerical and categorical attributes, including the immutability of features (e. g. gender, race) ; iv) is easily extended to other data such as images.